Blendpipe homeSend a brief
Menu

AI-assisted production with human review

Six tasks where AI-assisted tools help, each reviewed by a person and logged in the provenance log.

Send a brief
Used for
Concepting, previs, mattes, cleanup, upscaling, drafts
Not used for
A real person's face or voice without written consent
Every step
Reviewed by a person and logged
Delivered
Provenance log with a disclosure decision per step
A wooden weaving loom with a half-finished striped blanket, where a hand cropped at the wrist guides the threads and the shuttle passing through them is drawn in steel blue.
Live action
3D
Animation
AI-assisted

Shot 1, 0:00 to 0:04

Ridge at dawn

  • Live action
  • AI-assisted

A runner crests a frosted ridge as the sun clears the horizon.

Why this medium: Only a real location sells cold air and early light. An AI-assisted cleanup removes a power line from the sky; a compositor reviews every frame.

Fictional example. Select a shot number to read why it uses that medium. Orlenna is invented, not a client.

AI-assisted tools have specific jobs in a Blendpipe production: repetitive work, or helping people see an idea earlier. A person reviews every AI-assisted output before it is used, and every step is listed in the provenance log delivered with the files. This page describes the work. The policy behind it (consent, training data, disclosure and copyright) is on AI practices, and it is not repeated here.

The blend map above lights the AI-assisted track of the fictional Orlenna spot. In shot 1 (0 to 4 seconds) an AI-assisted cleanup removes a power line from the dawn sky, and a compositor reviews it frame by frame. In shot 5 (17 to 20 seconds) AI-assisted rotoscoping drafts the matte around the runner’s hand, and a person finishes the edges. That is 7 of the 30 seconds. The track shows only steps that touch a delivered frame: the AI-assisted previs of the ridge shot, made during development, is in the provenance log too, but none of it appears in the ad.

Six tasks, each with a review point and a disclosure outcome

The outcomes below follow the platform rules as checked on November 22, 2026. Each decision is recorded in the log, and your counsel makes the final call on anything unclear.

Concepting

  • What the tool does: produces reference frames for mood, lighting or composition during the concept and treatment.
  • Review point: the creative lead selects frames, labels each as AI-assisted inside the plan and keeps it out of final artwork. See creative development.
  • Disclosure outcome: planning only; nothing appears in the delivered ad. YouTube’s rule lists production help, such as scripts and ideas, among uses that do not need disclosure.

Previs

  • What the tool does: generates a rough moving version of a shot to plan camera, lens and timing before a shoot or render.
  • Review point: the director and producer check it against the board; it is replaced by the real shot.
  • Disclosure outcome: planning only, logged, not in the delivered ad.

Rotoscoping

  • What the tool does: drafts mattes that separate a person or object from the background, frame by frame.
  • Review point: a compositor checks every frame and finishes the edges by hand.
  • Disclosure outcome: a matte only separates layers of footage that was filmed; it does not add a person, place or event. The log records the shot and the decision.

Cleanup

  • What the tool does: removes unwanted objects such as a power line, a rig, a sign or a tripod shadow.
  • Review point: a compositor reviews every frame of the shot.
  • Disclosure outcome: assessed shot by shot. YouTube asks creators to disclose realistic altered footage of real events or places, and does not require it for minor edits. The log records whether the change alters what a place or event looks like, and which way the decision went.

Upscaling

  • What the tool does: raises the resolution of an archive clip, a render or a plate.
  • Review point: checked at full resolution for invented detail, especially in text, logos, faces and product edges.
  • Disclosure outcome: Meta’s and Google’s political and election ad rules list resizing and color correction as exempt changes. An upscale that invents visible detail is assessed like any other alteration.

Localization drafts

  • What the tool does: drafts transcripts, caption files and subtitle translations for language versions.
  • Review point: a fluent person checks every line against the final audio, the brand’s terms and the market’s legal lines. See campaign versioning.
  • Disclosure outcome: reviewed text drafts do not change what anyone in the ad appears to say. A synthetic voice in a language version is a different matter: it is not produced without the written consent of the person whose voice it imitates, and it is labeled under each platform’s rule.

What AI-assisted tools are not used for

Without written consent, no real person’s face, body or voice is generated, altered or cloned. Synthetic performers and digital replicas are used only if the client asks for them in writing, and then the plan includes the disclosures that apply. In New York, for example, S8420-A (signed December 11, 2025, in effect from the 180th day after, which we count as June 9, 2026) requires producers of commercial ads to “conspicuously disclose” a synthetic performer when they have actual knowledge of it, with penalties of $1,000 for a first violation and $5,000 after. TikTok also requires labels on realistic AI-generated or significantly edited content, and does not allow the likeness of private people without consent. Union terms, consent rules and training data are covered on AI practices.

The AI use entries in the provenance log

Every AI-assisted step gets a row in the provenance log, which is part of the handover described on deliverables. Each row records:

  • the step and the shot (with its frame range) it touched;
  • the tool category, such as “AI-assisted rotoscoping” or “generative image model”, with the tool’s public terms checked for commercial use and training on inputs;
  • whether the output appears in a delivered frame or only in planning;
  • the person who reviewed it, and the date;
  • the disclosure decision and the platform rule it followed;
  • the Content Credentials status: kept, added, or lost, with the reason.

That last column matters: the C2PA explainer says provenance metadata can be removed, and that Content Credentials do not judge whether content is true. YouTube may label content carrying C2PA metadata automatically, and Meta’s labels rely on industry standard indicators as well as self-disclosure, so the log tells you what each file carries before upload.

Example (fictional). The Orlenna provenance log has three AI use entries.

Step Shot In final frame Reviewer Disclosure decision
Previs of the ridge at dawn 1 (planning) No Director None; planning only
Cleanup: power line removed from the sky 1 (0 to 4 s) Yes Compositor, every frame Assessed per platform and recorded
Rotoscoping: hand matte 5 (17 to 20 s) Yes Compositor, edges finished by hand Assessed per platform and recorded

What you send and what you get

You send any rules your brand or legal team already has on AI use, the platforms and markets the ad will run in, and written consent from anyone whose likeness or voice is involved. You receive the AI disclosure plan inside the production plan, review points for every AI-assisted step in the schedule, and the provenance log at handover. The AI disclosure checklist lists the questions to answer before launch.

Want to know exactly where AI-assisted tools would touch your campaign before you commit? Send a brief, and the production plan lists every AI-assisted step with its reviewer.